Ji Qi

Pu Li, Jiawen Qi, Qinyu Chen

Deploying large language models (LLMs) on mobile devices increasingly relies on heterogeneous execution, yet no prior study has systematically characterized NPU effectiveness at the operator and pipeline level. We present the first stage-aware, multi-level benchmarking study of mobile LLM inference on a CPU-NPU heterogeneous SoC. We introduce an OPMASK-based controlled pipeline decomposition methodology that isolates communication, quantization, and computation overheads within the NPU execution path. Our results reveal a counter-intuitive stage-level performance reversal: CPUs outperform NPUs in the compute-intensive Prefill stage (up to 1.6x), while NPUs provide only limited acceleration in the memory-bound Decode stage (1.05-1.2x). We further show that scheduling overhead and cross-backend fallback reduce the practical benefits of NPU offloading. For the energy trend, increasing NPU offloading leads to higher energy consumption (up to 51%). Based on these findings, we derive design guidelines for NPU architects targeting on-device LLM inference.

Tianyu Yu, Zefan Wang, Chongyi Wang et al. · tsinghua

Multimodal Large Language Models (MLLMs) are undergoing rapid progress and represent the frontier of AI development. However, their training and inference efficiency have emerged as a core bottleneck in making MLLMs more accessible and scalable. To address the challenges, we present MiniCPM-V 4.5, an 8B parameter model designed for high efficiency and strong performance. We introduce three core improvements in model architecture, data strategy and training method: a unified 3D-Resampler model architecture for highly compact encoding over images and videos, a unified learning paradigm for document knowledge and text recognition without heavy data engineering, and a hybrid reinforcement learning strategy for proficiency in both short and long reasoning modes. Comprehensive experimental results in OpenCompass evaluation show that MiniCPM-V 4.5 surpasses widely used proprietary models such as GPT-4o-latest, and significantly larger open-source models such as Qwen2.5-VL 72B. Notably, the strong performance is achieved with remarkable efficiency. For example, on the widely adopted VideoMME benchmark, MiniCPM-V 4.5 achieves state-of-the-art performance among models under 30B size, using just 46.7\% GPU memory cost and 8.7\% inference time of Qwen2.5-VL 7B.

Qinyu Chen, Jiawen Qi

Vision-Language Models (VLMs) deliver impressive performance in understanding visual content with language instructions. However, redundancy in vision tokens results in the degenerated inference efficiency of VLMs, which hinders real-time use on edge consumer devices such as AR/VR devices. Existing efficiency methods commonly prune visual tokens using learned saliency, sparse attention schedules, or controller policies, but they often require architectural modification or access to intermediate activations. These pipelines add inference-time modules that increase compute and memory and often lead to an accuracy trade-off. Moreover, they also suffer from misalignment between the prompts and the region of interest in the images. Without human guidance, the model may focus on the wrong regions and miss small, high-frequency details when prompts or scenes change. In this paper, we propose GazeVLM, a training-free framework that uses the human eye gaze as a natural supervisory signal to allocate computation where it matters. By extracting gaze-driven regions of interest (ROIs) and optionally combining them with a low-resolution global view, GazeVLM mimics fovea-periphery perception to cut redundant visual tokens while preserving task-relevant details. We evaluate the visual question answering tasks on Qwen2.5-VL-3B/7B on the VOILA-COCO benchmark with human gaze. Quality of the answer is assessed by GPT-4o pairwise judging and a weighted score over coverage, accuracy, details, and fluency. Efficiency is measured by token counts and FLOPs. GazeVLM reduces visual tokens by up to 93.1%, total tokens by up to 59.6%, and FLOPs by 50%, while keeping better answer quality relative to full-resolution baselines. Our results show that aligning model computation with human gaze offers a simple, plug-and-play path toward efficient VLM inference on consumer devices.